Media Buying in the Andromeda Era

Search for Andromeda and you will find forty confident rules and one engineering post. The rules usually come with numbers: this many ads, that much budget, a similarity score nobody can show you.

I want to do this differently. In this playbook I will keep three things apart: what Meta has actually published, what operators reasonably infer from it, and what I do in accounts because it works. When I am guessing, I will say so. When I am quoting Meta, you will see a footnote.

What do we actually know about Andromeda and GEM?

Less than the internet claims, but enough to plan around.

Andromeda: the first cut

Meta says retrieval is the first step in its multi-stage ads recommendation system, narrowing tens of millions of ad candidates into a few thousand relevant ones for later ranking models.¹ It reported a 6% recall improvement and an 8% ads quality improvement on selected segments, and described a 10,000x increase in model capacity.¹ It also said that more than a million advertisers had used its generative AI tools to create more than 15 million ads in a single month.¹

The practical meaning: the pool of ads is growing, and the first filter has to be fast. My earlier first look at this is in Meta Andromeda: What the New Ad Retrieval Engine Means for Small Accounts.

GEM: the teacher behind ranking

Meta calls GEM its most advanced ads foundation model, built on an LLM-inspired approach and trained across thousands of GPUs. It transfers what it learns to hundreds of smaller models that serve ads.² Meta says it was 4x more efficient at driving ad performance gains than previous ranking models for a given amount of data and compute, and that GEM models behavior sequences of up to thousands of events.²

What Meta said at the start of 2026

Per PPC Land's report on Meta's Q1 2026 call, more than 8 million advertisers now use at least one of Meta's AI ad creative tools, up from 4 million at the end of 2024, mostly small and medium businesses. The same report says Lattice and GEM improvements drove a conversion rate increase above 6% for landing page view ads.⁴ On the quarter, Meta reported revenue of $56.31 billion, up 33%, with ad impressions up 19% and average price per ad up 12%.³

Read those as direction, not promise. They are platform-wide results from the platform itself.

What has Meta not told us?

Almost everything an advertiser would want in a rulebook.

As I read the two engineering posts, neither gives account setup instructions, an ideal number of ads, or a threshold for how different two creatives must be. GEM's post does not explain how it relates to Andromeda at all.² It does describe a future of "agentic, insight-driven advertiser automation."² That is a stated plan, not a shipped feature.

So here is my rule for the whole piece. If someone gives you a precise number of ads, a precise similarity score or a precise budget threshold and cannot point to a Meta page, treat it as one operator's opinion, mine included.

Layer one: how much creative variety do you really need?

Enough that the system has different reasons to show you to different people.

This is the part that follows most directly from the retrieval design. If a few thousand ads survive the first cut out of tens of millions, an ad that is nearly identical to another in your own account adds little. That is inference, and it matches what I see.

Two well-known Meta educators disagree on volume. Ben Heath said this in a video posted on November 11, 2025, which includes a Hyros sponsor read:⁵

Jon Loomer answered from the other side in a video posted on December 3, 2025: "More often than not, more ads means worse ads."⁶

I side with Loomer on volume and with both on variety. Concepts matter, copies do not.

Concepts, not edits

A concept is a distinct reason to stop scrolling: a different problem, promise, proof, person or format. An edit is the same reason in a new outfit. I use five levers when I build a concept map:

• Problem: the pain or situation the buyer is already in.
• Promise: the outcome you are offering.
• Proof: the evidence, whether results, a demonstration or a reason to believe.
• Person: who is speaking, a founder, a customer, an expert.
• Format: talking head, screen recording, text-led static, a before and after.

Take one offer and change one lever at a time until you have five or six ads that would not be mistaken for each other. For the testing mechanics, see Creative Is the New Targeting and the creative bank approach.

Where AI tools help and where they stop

Generative tools are good at producing formats and variations quickly. They are not good at choosing the problem your buyer cares about. If you let a tool make fifteen variants from one prompt, you have made fifteen edits very fast. Put the thinking into the concept map first, then use the tool to execute it.

Layer two: how should you structure the account?

Simpler than you would have built it three years ago. This layer is my practice, not a Meta requirement.

Audience slicing used to be the main lever. In an account where the delivery system does the matching, I would rather give each campaign real budget and real variety than split money across many small ad sets with near copies. In practice I organize by offer and funnel purpose, not by dozens of audience carve-outs.

There are exceptions. Different offers, different countries and regulated categories may need separation, and some compliance rules must be applied by hand. The point is to separate for a reason, not by habit. My notes on account structure go deeper on this.

Layer three: what signal are you sending?

This is the layer I would fix first, because automation scales whatever you teach it.

Meta says GEM learns from long sequences of user behavior and passes that learning to many ranking models.² My inference is straightforward: those models predict value from the outcomes the account reports. If the outcomes are an unfinished page view, they learn to find people who view pages. If the outcomes are qualified bookings and paid customers, they have something worth finding.

Why server-side signal and match quality matter

An event only helps if it is real, arrives promptly and can be matched to the right person. Browser-only tracking loses events to blockers and privacy settings. Events that are late, duplicated or unmatched teach the model about the wrong people, and a more automated system turns that flaw into a faster mistake. A sound server-side connection and strong match quality are how you protect the quality of what the model hears. For the basics, start with Meta Conversions API in 2024. In the system I build for clients I treat the signal path as the foundation, and I keep the details of my own implementation to myself.

Pick the deepest event you can feed

Optimize for the furthest-down outcome that still happens often enough to learn from. If paid customers are too rare, optimize for a qualified booking. Reaching for a shallow event because it is plentiful is how accounts end up with cheap, useless volume.

How do you test without confusing the system?

On a schedule, with rules written before the results arrive.

How do you know it worked?

Check the dashboard against cash. Meta changed how it counts attribution twice in the past year, so platform numbers can shift without your business changing. I explain the reconciliation in Cash ROAS vs Platform ROAS, and the case for testing causation in incremental attribution.

What mistakes do I see most often?

Five, and none of them is exotic. Everything in this list is my practice and my observation, not Meta guidance.

• Counting ads instead of ideas. Twenty ads built from one script is one test, not twenty.
• Judging too early. A new concept needs enough spend and enough time for the downstream outcome to show up. Cheap clicks are not a verdict.
• Rebuilding on a bad week. Every restructure resets what the account has learned. Change the account on a schedule, not on a mood.
• Feeding the system a shallow event. If the reported outcome is a page view, the account will find page viewers all day.
• Skipping compliance review in regulated categories. If you advertise health offers, every new concept is a new thing to review, so build that review into the batch before launch. My notes on Meta's health and wellness restrictions cover the data side.

The common thread is impatience. Variety needs a schedule: a batch of different ideas, a fixed window to judge them and a rule you wrote before you saw the numbers.

Your first 30 days

Do less, in the right order.

Week one, audit your signal and fix the worst event problem. Week two, build the concept map and brief five genuinely different concepts. Week three, launch them as one batch inside a simplified structure and leave them alone. Week four, review against cash and the kill rule you wrote in advance.

The platforms keep getting more automated, and that makes your inputs matter more. Give the system distinct ideas and honest outcomes, and your inputs become the edge. Give it near copies and noisy events and it will faithfully scale both.

Keep building.Ekai